Hi everybody,

We are searching a good OS for a cluster which will eventually scale thousands of processors. This cluster will be primarily for running MPI-based codes on a Myrinet MX and higher connection system.

To ease the task of our technical support we are thinking about linux-based OS on bootable CD or DVD like ClusterKnoppix or ParallelKnoppix such that no installation will be required and an easy upgrade is possible.

What is your opinion?

Thanks a lot in advance.

> What is your opinion?

Depends on the rest of the decisions because the OS is not important on the nodes of a computational cluster. The first 4 parallel computers I used had no OS running on the nodes at all.

Red Hat (i.e. RHEL, Centos or ScientificLinux) is the most common OS running on larger computational clusters and therefore is probably the easiest to support for large clusters.

If you are new to clustering then going with a widely used cluster project might be wise. I would not recommend going with small or dead cluster projects. ClusterKnoppix looks dead and is based on OpenMosix which has been end-of-lifed. ParallelKnoppix is alive but small and likely to die when whoever is behind it loses interest.

regarding to diskless cluster, look at: http://www.clustermatic.org/

Try the following;

http://www.rocksclusters.org

http://oscar.openclustergroup.org/

http://www.yaci.org/

Thank you very much to all of you!

>The first 4 parallel computers I used had no OS running on the nodes at all.

Really? That sounds interesting. I am aware of a variety of options for diskless clusters, loading the OS from eprom, floppy, or CD to RAM, but no OS at all is news to me. How does the master machine control the hardware (processor, network interface, memory a.s.o) on the node without any OS residing in the node's RAM?

> How does the master machine control the hardware (processor, network
: interface, memory a.s.o) on the node without any OS residing in the node's
: RAM?

It doesn't. It simply sends to messages to the nodes via whatever hardware is in being used. The nodes in turn simply send message to their neighbours via whatever hardware is in use.

On each node all that runs is a single copy of the slave program writen in C, Fortran or whatever. The small amount of hardware driver code that would usually be in the kernel is linked to the slave program.